[alexandre-vl]

Closes #1001.

Forking a chat while one of its subagents ran left that subagent spinning in the fork forever. The fork copies the source's entries as they are, so the spawn chip came across running, and no run in the fork will ever settle it. The copy now settles it failed, as the engine already does for a running chip it can no longer follow: at run end (crates/engine/src/sessions.rs:4824), and in the restart sweep, which settles running chips in completed parent turns too (mark_abandoned_streams, crates/engine/src/doc_host.rs:949-973). The source keeps its own chip and its real outcome.

What changes

• Fork. ForkSideChat already resolved the copied Input parts. It now also settles each copied tool part whose subagent is Running (crates/engine/src/rpc.rs:2753-2763). The subagentRef, settled chips and the source are left alone.

Testing

• macOS 27.0, rustc 1.99.0: cargo test --locked -p zeron-engine --lib --test side_chats --test restart_resume --test e2e gives 630 passed (5 ignored), 7, 11 (1 ignored) and 29 (2 ignored).
• CI now runs side_chats: it is named right after restart_resume in the engine's list of integration binaries (scripts/ci/test-core.sh:27), which didn't include it. Providers: a switch to turn native subagents off #1046 adds the same change, so the two merge in either order.
• a_fork_settles_the_running_subagent_chips_it_copies (crates/engine/tests/side_chats.rs:405, next to the six other ForkSideChat tests) forks one entry holding a running chip, a done chip and an open question. The fork gets (main--sub--running, failed), (main--sub--done, done) and an answered question; the source keeps running, done and its open question. It fails at side_chats.rs:510 with the fix reverted (left: ([(Some("main--sub--running"), Some(Running)), …], [true])), with a copy that drops the ref (left: ([(None, Some(Failed)), …], [true])), and with the question left open (left: (…, [false])).
• Live, Claude Code 2.1.310 with Haiku, on headless engines with their own data dirs. A chat starts one background Agent whose prompt asks for sleep 90 in the foreground, ends its turn, and is forked while that subagent still runs: the source's chip read running just before the fork in both runs. Claude Code's Bash tool refused the sleep 90 as a standalone sleep, so the subagent ended on its own, after the fork.
  • v0.2.122: the fork's chip read running right after the fork and still did 146 s later, while the source's had gone done.
  • This branch: the fork's chip read failed, its subagentRef intact, from the first read after ForkSideChat returned to the last, 149 s later. The source's went from running to done. The subagent doc behind that ref still serves the subagent's transcript, which is what the red chip opens when the transcript blob isn't there (crates/ui/src/shell.rs:4624).
• Rebased onto v0.2.124 (c04e3749) after these runs, with no conflict. Merged there with the other five PRs of this series (Claude subagents: count and open only this chat's spawns #1041, Claude subagents: forward a foreground subagent's text and thinking #1042, Claude subagents: give a typed /skill's fork a chip and a tab #1043, Claude subagents: fill a typed fork's tab while the CLI holds it #1044, Providers: a switch to turn native subagents off #1046), it passes every suite above again, plus the rest of CI's selection, sync's mock-server tests, zeron-mcp and the UI's settings:: tests, each executable checked to be built from that merge. The one failure, preview's backgrounded_scrubbed_terminal_and_out_of_tree_servers_stay_attributed, fails the same way on v0.2.124 alone, on this Mac.
  • Live on that merged build, the same fork: its chip read failed in all 30 samples over 149 s, and the source's went from running to done.
• Not run: Windows and Linux, the full app by hand.

Known limits

• In the fork the chip turns red like a failed subagent, even when the subagent finishes fine in the source. A fork never sees that outcome, and failed is what the doc already uses for a subagent cut off mid-run.
• The status is display-only. A click on the chip still opens the subagent's real transcript, from the source's namespace (crates/ui/src/shell.rs:4582-4587), and the fork's history prompt leaves the status out (crates/engine/src/sessions.rs:2723).
• Forks made before this change keep spinning: the restart sweep only visits chats with a stale journal (recover_stale, crates/engine/src/sessions.rs:1516-1545), and a fork that never ran has none.
• Questions: show what was answered, and close one without answering #1022 rewrites the same Input branch of the copy: a question still open in the source gets empty answers in the fork. The two conflict as text only. Whichever lands second moves Questions: show what was answered, and close one without answering #1022's if !*resolved { *answers = Some(Vec::new()); } into this match's Input arm, next to *resolved = true. They change different parts, and this PR's test only reads resolved.
• The engine's CI list is one line (scripts/ci/test-core.sh:27). Claude subagents: give a typed /skill's fork a chip and a tab #1043 adds claude_subagents to it, and Claude: show Workflow runs as subagents with live progress #784 (with Claude: open each Workflow agent's live transcript #785 on top) workflow_progress, so each conflicts with this PR there. Whichever lands second keeps every name.

Screenshots

Offscreen renders of the production Shell (Window::render_to_image), attached to each engine; docs/media/fork-settles-subagents/README.md says how they were made. They come from an earlier run than the one under Testing: v0.2.121, with a subagent running sleep 70. The media are in their own commit, 54e5490f, if you'd rather drop them before merging.

v0.2.121, an hour after the fork	This branch

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}

[zeronsh]

Security audit clean: engine-only change to ForkSideChat (settles copied running spawn chips failed, matching run end and the restart sweep), one test, docs, media PNGs (no trailing data), and side_chats added to the CI test list. Rebased onto main; locally on Linux: side_chats 7/7, restart_resume 11, e2e 29 pass; the new test fails with the fix reverted; no new clippy warnings in touched files. Reflowed the docs/mcp.md paragraph. Thanks!